PIXELOKAY / AUDIO TOOLS NO PAID API

AI Speech Cleanup Editor

Find long pauses, spot filler words and repeated phrases, and make precise cuts without uploading your recording. Review every suggestion, edit timed words and export a cleaned WAV with the pauses removed.

✓ Original never overwritten✓ Review every cut✓ Download WAV + transcript✓ Private by default

UPLOAD · REVIEW · CUT · EXPORT

Speech cleanup studio

Free · Browser based

Add your recording

Audio remains on this device.

Choose audio or drop it hereWAV, MP3, M4A, OGG and other browser-decodable audio · up to 150 MB

The demo contains tones and pauses, not speech. Use the bundled spoken sample for filler-word testing.

No audio loaded.

Timed transcript

Import or transcribe to find spoken fillers.

Import transcript (optional)

For exact word cuts, supply CSV/JSON word timings, or use experimental on-device transcription. SRT/VTT cues can be reviewed but are not treated as word-aligned.

Reopen saved edit plan (.json)

Reopen a project only after loading the same original audio. Where Web Crypto is available, audio fingerprints protect against applying cuts to another recording. Otherwise the filename and duration must match; review carefully.

Model not loaded

Experimental transcription is limited to 90 seconds. Whisper Tiny English downloads free model files on first use (network required). Transcription and word timing are approximate. No audio upload. The model downloads only after you select Download & Load in the setup popup. Transcribe opens setup if needed; it never starts an automatic download.

Detection settings

Control the cleanup without guesswork.

Silence threshold

Minimum pause

Leave a pause

Cut crossfade

Suggestions are not destructive. Deselect any cut before exporting. Stronger silence settings can remove quiet speech; listen to the result.

Make a precise cut

Optional manual trim from original timeline.

From (seconds)

To (seconds)

The waveform is clickable: click once to seek, then use playback to find accurate cut boundaries.

Preview and export

Every cut remains reviewable.

Export filename

Timed subtitles

WAV is 16-bit PCM, mono/stereo, at the decoded sample rate (which can differ from the source). The JSON edit plan stores timestamps—not the recording. Saved edit plans can be reopened with their matching original audio. Captions include only fully retained timed entries; partially cut words are omitted, and TXT marks them for review. Always listen and verify timings.

SOURCE TIMELINE

Your waveform will appear here

0:000 cuts0.0 s removed

0:000:00

Audio level Proposed cuts Playhead

The blue waveform shows the original recording. Rose overlays show proposed removals. All timestamps refer to the original audio.

REVIEW BEFORE CUTTING

Cleanup suggestions

Load audio and click Analyze recording to find pauses and potential words to remove.

No suggestions yet. Your original audio stays untouched.

EDIT WORDS · VERIFY TIMINGS

Editable transcript

0 entries

Edit speech text and adjust start/end times in seconds. Edits clear stale automatic suggestions. For cue-level imports, use manual cuts or silence suggestions for precision. Transcript edits do not regenerate the spoken audio.

Start	End	Speech	Action
Import a timed transcript or transcribe locally.

0 timed entries

KNOW YOUR EDITS

How speech cleanup works

01 / Analyze

Find quiet gaps accurately

Silence detection uses measured signal energy. Pauses near timed words are excluded, so proposed cuts do not automatically erase speech detected by the transcript.

02 / Review

Keep editorial control

Filler words are proposed when reliable word timestamps exist. Repetitions and false starts are suggestions, not automatic deletions. Listen and decide.

03 / Export

Get editable decisions

Download an edited WAV, timed SRT/VTT captions, text transcript and a reopenable JSON edit decision list. Original files remain unchanged. No paid API or audio-upload backend is involved.

Frequently asked questions

Does this automatically transcribe every recording?

No. Optional local Whisper transcription depends on browser compatibility, first-time model downloads, device memory and language. Importing an existing timed transcript is always available.

Can SRT subtitles be used for removing filler words?

SRT/VTT contain cue timestamps, usually covering sentences. The editor imports them but does not assume exact word timings. You can still make manual cuts and detect true quiet pauses.

Will cuts sound completely seamless?

Short configurable crossfades help soften abrupt joins. They do not guarantee natural phrasing, especially within a word or in noisy recordings. Preview your edits before exporting.

Are recordings sent to a cloud service?

No. Decoding, silence detection and WAV rendering occur in this browser. Optional Whisper downloads model files from a public host, but sends no audio to that host.

Using AI Speech Cleanup Editor

Import a recording and optional timed transcript or request experimental transcription. Analyze, review every suggested removal and adjust manual cuts before rendering and listening. Save the edit plan with the matching original audio.

Offline energy analysis is conventional; experimental ASR downloads a model/runtime and supports shorter recordings. Cue-level subtitles are not word timing and cannot justify precise filler cuts. Caption export omits partly cut entries rather than inventing timing. Sampled fingerprints are not authentication; fallback matching can use filename/duration with warnings. Projects contain no audio or autosaved drafts. Resource limits and browser decoding apply.

Review transcript and silence suggestions before cutting audio

Import a recording and optionally captions or Whisper text, analyze quiet intervals and word-timed suggestions, then make manual edits and preview the cuts. The editor does not overwrite the original recording; export the rendered WAV, captions, transcript or edit plan you need.

Example workflow

Keep the original audio and load a supported recording; add captions or request the optional short-recording transcription if useful.

Review each suggested cut against the waveform and audio. Treat sentence-level SRT/VTT cues as broad references, not word-level timing.

Adjust cut boundaries and crossfades, preview the result, then export WAV and any transcript/caption files. Reopen an edit plan only with the matching original recording.

Authored example: In a spoken interview, inspect a long pause that follows a guest's answer. Keep a quiet breath or word onset if removing it changes the phrasing, then replay the join before rendering.

Limits and model handling: Quiet-interval detection is conventional; optional English Whisper downloads runtime/model assets after explicit transcription setup and supports shorter recordings. Crossfades can soften joins but do not guarantee natural phrasing. Partly cut SRT/VTT cues are omitted rather than assigned invented times. Edit-plan JSON contains no audio or autosaved draft; sampled fingerprints are not authentication and resource/browser-decoding limits apply.

AI Speech Cleanup Editor questions

Can sentence-level captions identify exact filler-word cuts?

No. SRT/VTT cues often span sentences and are not word timings. Review the audio and set precise cuts manually.

Does the saved edit plan include the recording?

No. The plan must be used with the matching original audio. Keep that source file separately.